University of Bologna, A.Y. 2025–2026.
edoardo.merli2@unibo.it
🚨You can submit this assignment until January 31st 2027. After this date, you must complete the 2026–2027 A.Y. assignment instead.🚨
Over the last few years, the field of Person Re-Identification (Re-ID) has witnessed significant progress, achieving impressive performance on standard benchmarks. However, traditional Re-ID methods typically operate on pre-cropped images, assuming that pedestrians have already been perfectly localized by a separate detector.
This assumption rarely holds in real-world applications, where the system is not fed perfect crops. In these scenarios, it must ingest full scene images, locate all individuals, and identify the target person simultaneously, handling challenges like background clutter, occlusion, and pose variations. Addressing this gap is crucial for practical intelligent surveillance systems, such as finding a missing child in a crowded park or tracking a suspect across multiple unlinked cameras.
This task, which aims to jointly solve Pedestrian Detection and Person Re-Identification, is known in the literature as Person Search. Unlike standard Re-ID, where the query is matched against a gallery of cropped bodies, Person Search requires matching a query person against a gallery of raw, uncropped scene images, making it a more challenging and realistic "in the wild" problem.
Your goal is to implement a model that is able to tackle the Person Search task, which can be seen as the combination of pedestrian detection and person re-identification: given a query image with a single bounding box enclosing a person, detect and match that same person in the other images of the dataset, based on body shape and clothes.
Here are a couple of examples of query image and other images with the same person detected and matched.
Test set images and identities should remain unseen during training and only be used at test time.
The dataset you are going to work with is called "Person Re-identification in the Wild" (PRW) [1], which was proposed for the task of Person Search.
The dataset contains 11816 images annotated with 43110 pedestrian bounding boxes, among which 34304 are linked to an identity (ID randing from 1 to 932) while the rest are not (ID = -2, corresponding to "ambiguous person").
It is divided into a training set with 5704 images and 482 IDs and a test set with 6112 images and 450 IDs. The training set is further split into 5134 and 570 images, both spanning the same 482 IDs.
A summary of the dataset statistics is provided below.
| Datasets | #frame | #ID | #ped. | #ped. w/ ID | #ped. w/o ID |
|---|---|---|---|---|---|
| Train | 5134 | 482 | 16243 | 13416 | 2827 |
| Val. | 570 | 482 | 1805 | 1491 | 314 |
| Test | 6112 | 450 | 25062 | 19127 | 5935 |
| Total | 11816 | 932 | 43110 | 34034 | 9076 |
The dataset is available at this link.
In order to achieve the aforementioned goal, you are allowed to use any method of your choice, either existing (see References), re-adapted, or even invented by yourself. From an implementation standpoint, you can:
- Use off-the-shelf pre-implemented PyTorch models.
- Use pre-trained weights.
- Use portions of existing code found in public repositories. If so, cite the source you took the code from.
- Use any existing library for computing metrics, displaying results or providing any other functionality you need.
Teams are not allowed: the assignment must be done individually.
You must submit the following files:
- A notebook called
main.ipynb. You can either put all the code into the notebook or separate it into python scripts referenced by the notebook and use the notebook only for visualization purposes and textual explanations. The notebook must be runnable: disable the training (if you leave the training code inside the notebook) and load the model(s) weights to run inference. The first cell of the notebook must contain your student ID, full name, and institutional email. - The weights of your model(s) saved as
.ptfile(s). - A
README.mddescribing your project structure and providing instructions on how to run it: specifically, tell me if I should run it locally or on Colab, Kaggle, etc. - If the notebook runs locally, a
requirements.txtcontaining the packages I need to install to run it. - Any other file needed to run the notebook.
If you use Kaggle, you can choose to upload the model weights to Kaggle and add them to the Kaggle notebook, like we did in the hands-on sessions. Make sure they are public in this case.
(Extra: if you choose to use Kaggle and feel brave enough, you can decide to design your code to run using both T4 GPUs via distributed training: this Kaggle notebook explains how to do so. This can be a great skill to have in your toolset, regardless of the assignment.)
You must submit your files through the dedicated assignment submission form in the course page on Virtuale. You can submit at most 20 files with maximum size 20 MB each (both limits imposed by Virtuale itself). In case some of your files exceed 20 MB (e.g. the model weights), upload them on your institutional OneDrive and add a public link to them in your README.md.
Make sure to submit the assignment before you write to prof. Salti to schedule the oral exam. Put me in cc when you do so, so that I will see when your exam gets scheduled and I can make sure your assignment has been graded by the time you attend the exam. The assignment can only be submitted once per academic year.
My evaluation of your work will be based on:
- Qualitative results, i.e. ranking of the first few detected and identified people with respect to the query image and person (from the test set). An example of qualitative is the image above.
- Quantitative results (a.k.a. metrics), i.e. mAP and top-1 ranking. These metrics should guide you in the design and tuning of your model. Look into papers showing experiments on PRW (e.g. [2, 3, 4, 5]) to know the range of values you should strive for. N.B.: a detected box is considered a true positive if it has an IoU > 0.5 with the ground truth box and the ID matches.
Important
In order to compute metrics in a way that is consistent with the literature and uniform among you, make sure to adapt the eval_search_prw function I have provided you (here) to work with your code. Intuition for the mAP in ranking tasks can be found here.
- Experimental procedure. Do not just show me the results of your final model; instead, perform ablation studies that (empirically) prove how your main choices positively affect the model performance. In a practical sense: you are expected to experiment with variations (not many, 1 or 2 are enough) of your model and compare how they change the model performance. Examples of valid variations are: architecture modification, change of pretrained model backbone, change of loss function, etc.; an hyperparameter change (e.g. learning rate) is not a valid variation of your model.
Any result you show must be critically discussed and/or interpreted and have an actual impact on your investigation. - Clarity of exposition and code readability.
main.ipynbmust be a combination of textual and code cells whose content is intertwined: do not write text cells referencing code that I cannot see anywhere (not even in external scripts) nor write non-trivial code cells with no textual explanation preceding them. External scripts, if present, should be well-structured (classes, functions, etc.), well-written (mnemonic variable names, no random commented lines, etc.) and well-documented. Plots and tables are always welcomed.
Keep in mind that this is not a challenge: you will not be ranked based on the metric scores you achieve. Therefore, points 3 and 4 will be as impactful on the final grade as points 1 and 2.
You will receive a score from 0 to 10, that will be summed to your oral exam score yielding a total of ≤ 32 (= 30 + laude) points. The assignment grade will be displayed on Virtuale before you attend the oral exam.
Warning
The score received for points 1 and 2 is based exclusively on the qualitative and quantitative results displayed in the notebook: no further code will be written by me to investigate performance or visualze outputs. Therefore, make sure to clearly show performance metrics and plot a sufficient amount of qualitative outputs.
What follows is a (non-exhaustive) list of papers you might want to take a look at:
[1] Zheng et al. "Person Re-identification in the Wild". CVPR 2017.
[2] Jaffe et al. "Gallery Filter Network for Person Search". WACV 2023.
[3] Yu et al. "Cascade Transformers for End-to-End Person Search". CVPR 2022.
[4] Dong et al. "Instance Guided Proposal Network for Person Search". CVPR 2020.
[5] Yan et al. "Anchor-Free Person Search". CVPR 2021.
This section will be dynamically updated in the following weeks/months as I receive questions about the assignment.
Q: Can I clone a Github repository and work with that code?
A: Your code should start from scratch and can integrate portions of existing repositories, for example the model architecture code, but cloning an entire Github repository would heavily penalize your assignment grade (and not make you learn as much).
Q: Should I build multiple models for a comparison in the end?
A: You are expected to experiment with variations (not many, 1 or 2 are enough) of your model and compare how they change the model performance as part of point 3 of the evaluation score breakdown "Experimental procedure".
Q: In the provided eval_search_prw function, should parameter ignore_cam_id be set to True (as default) or False?
A: The parameter ignore_cam_id makes the task harder if set to False by excluding gallery images captured by the same camera of the query. It should be set to True as the other works in the literature use this default setting (e.g. see COAT code), in order to have a fair comparison.
Q: Should the evaluation be performed using only the 2057 query images present in folder query_box or should we consider as queries all the people extracted from the detector to which the model assigns an identity?
A: The model should be evaluated using as queries the bounding boxes listed in file query_info.txt, which refer to the corresponding images in folder query_box.
Q: Some identities appear in both the training and test splits, but the assignment requires that "Test set images and identities should remain unseen during training". How should we handle this?
A: The dataset's README states: In detail, "frame_test.mat" and "frame_train.mat" specify the train/test frames, and "ID_test.mat" and "ID_train.mat" specify the train/test IDs. Note that a small portion of IDs used in training may appear in the testing frames, but will not appear in the testing IDs. The testing IDs are those used in the queries (query_box folder and query_info.txt) so there is no overlap in the Re-ID task. It's therefore not necessary to handle this in any particular way.
Q: I cannot find the validation set and there are a few discrepancies between the assignment table and the numbers extracted from the dataset. Is this normal?
A: The table is the one reported in the PRW paper, the dataset is the one published by the authors. Discrepancies arise because the work of people we rely on is not always perfect, and I wanted to give you the informations as they are reported online to get you used to handle these discrepancies by yourself, without requiring any kind of supervision. The missing validation set is another example of such discrepancies.